Gyrokinetic Dynamic Fidelity Refinement

Most gyrokinetic codes are designed to run on high-end supercomputers, including exascale devices. Key problems that such codes are used for include microstability, turbulence, transport, and concept optimization. As part of "whole device modeling" in particular, we need accurate calculations of these phenomena on teraflop devices (or even gigaflop) when the gyrokinetic phenomena comprise only part of a complex problem. Doubling the resolution of each dimension of a 5D + time problem requires an approximately 64X increase in resources at a minimum (for linear scaling of resources with problem size). Tripling the resolution in all dimensions with typical application performance available today (eg, to enable coupled ITG/ETG turbulence calculations) typically requires well over 1000X increase in resources. There is a major opportunity in these grim assessments, for algorithms that can dynamically choose the mesh discretization levels and even the physics model automatically, as demanded at any instance within any particular calculation, particularly to meet user-supplied uncertainty targets. We present key advances toward this goal of "dynamic fidelity refinement" (DFR). Our prototype electromagnetic, gyrokinetic DFR algorithm runs on a desktop.